Aging researcher Morgan Levine proposes training AI to recognize biological states to guide treatment selection

Aging researcher Morgan Levine proposes training AI to recognize biological states to guide treatment selection

In her November 6 essay, she proposes training models on biological observations and the results of interventions. A model would identify features in the measurements that predict how a cell, tissue, or patient will respond to treatment.

In October, Levine wrote that cell types and diagnoses can obscure important differences. She explains her current proposal using the example of catching a ball: we track its position and movement. That is enough to put a hand in its path; calculating the interactions of every molecule in the ball and the air would be unnecessary. Levine suggests that a limited set of features might also suffice to predict the response to a biological intervention.

Instruments measure the activity of thousands of genes simultaneously. Levine uses “perception” to describe a model’s ability to identify features in these measurements that predict how a cell or tissue will respond to environmental changes and interventions. The relevant features depend on the task: predicting cell division may require different features from predicting tissue repair after injury.

In drug discovery, scientists propose a hypothesis about the mechanism of a disease and select a molecular target, such as a protein whose activity they want a drug to alter. The effects on tissues and organs depend on that protein’s interactions with other components and on the state of the organism. Levine proposes starting with an individual patient’s state and the desired treatment outcome, then using a model to select an intervention that accounts for effects across tissues, benefits, and harms. The aim is to learn to “deliberately move biological systems from one state to another.” This could involve a single drug, a combination of drugs, or another type of intervention. To select a treatment, the model must account for cause and effect, even if it represents these relationships differently from the diagrams scientists use to describe molecular interactions.

In this approach, each experiment would both test a treatment and teach the model to distinguish biological states more accurately. If two states that appear similar to the model respond differently to an intervention, the prediction error could reveal a distinction it missed. Levine proposes selecting experiments that would help resolve uncertainty that impedes prediction or treatment selection. The results would help the model distinguish states more accurately and predict the effects of subsequent interventions. Levine proposes testing the reliability of these predictions on interventions, groups of people, and biological conditions the model has never encountered.

Levine also links reliable recognition of biological states to the possibility of shortening some clinical trials. AI could identify early changes in the body that predict the subsequent course of a disease or the benefit of treatment. If studies in humans confirm that an intervention-induced change in such a measure predicts patient benefit, researchers could decide sooner which treatments to continue testing.

Biotechnology entrepreneur Tyler Golato proposes letting AI improve research tools. “There is a risk in assuming that we are using the right instruments and analytical methods,” he writes. Levine endorsed this view.